Re: [v3 PATCH] mm: ksm: do not block on page lock when searching stable tree

From: John Hubbard
Date: Wed Jan 30 2019 - 02:14:25 EST


On 1/29/19 12:29 PM, Yang Shi wrote:
ksmd need search stable tree to look for the suitable KSM page, but the
KSM page might be locked for a while due to i.e. KSM page rmap walk.
Basically it is not a big deal since commit 2c653d0ee2ae
("ksm: introduce ksm_max_page_sharing per page deduplication limit"),
since max_page_sharing limits the number of shared KSM pages.

But it still sounds not worth waiting for the lock, the page can be skip,
then try to merge it in the next scan to avoid potential stall if its
content is still intact.

Introduce trylock mode to get_ksm_page() to not block on page lock, like
what try_to_merge_one_page() does. And, define three possible
operations (nolock, lock and trylock) as enum type to avoid stacking up
bools and make the code more readable.

Return -EBUSY if trylock fails, since NULL means not find suitable KSM
page, which is a valid case.

With the default max_page_sharing setting (256), there is almost no
observed change comparing lock vs trylock.

However, with ksm02 of LTP, the reduced ksmd full scan time can be
observed, which has set max_page_sharing to 786432. With lock version,
ksmd may tak 10s - 11s to run two full scans, with trylock version ksmd
may take 8s - 11s to run two full scans. And, the number of
pages_sharing and pages_to_scan keep same. Basically, this change has
no harm >
Cc: Hugh Dickins <hughd@xxxxxxxxxx>
Cc: Andrea Arcangeli <aarcange@xxxxxxxxxx>
Suggested-by: John Hubbard <jhubbard@xxxxxxxxxx>
Reviewed-by: Kirill Tkhai <ktkhai@xxxxxxxxxxxxx>
Signed-off-by: Yang Shi <yang.shi@xxxxxxxxxxxxxxxxx>
---
Hi folks,

This patch was with "mm: vmscan: skip KSM page in direct reclaim if priority
is low" in the initial submission. Then Hugh and Andrea pointed out commit
2c653d0ee2ae ("ksm: introduce ksm_max_page_sharing per page deduplication
limit") is good enough for limiting the number of shared KSM page to prevent
from softlock when walking ksm page rmap. This commit does solve the problem.
So, the series was dropped by Andrew from -mm tree.

However, I thought the second patch (this one) still sounds useful. So, I did
some test and resubmit it. The first version was reviewed by Krill Tkhai, so
I keep his Reviewed-by tag since there is no change to the patch except the
commit log.

So, would you please reconsider this patch?

v3: Use enum to define get_ksm_page operations (nolock, lock and trylock) per
John Hubbard
v2: Updated the commit log to reflect some test result and latest discussion

mm/ksm.c | 46 ++++++++++++++++++++++++++++++++++++----------
1 file changed, 36 insertions(+), 10 deletions(-)

diff --git a/mm/ksm.c b/mm/ksm.c
index 6c48ad1..5647bc1 100644
--- a/mm/ksm.c
+++ b/mm/ksm.c
@@ -667,6 +667,12 @@ static void remove_node_from_stable_tree(struct stable_node *stable_node)
free_stable_node(stable_node);
}
+enum get_ksm_page_flags {
+ GET_KSM_PAGE_NOLOCK,
+ GET_KSM_PAGE_LOCK,
+ GET_KSM_PAGE_TRYLOCK
+};
+
/*
* get_ksm_page: checks if the page indicated by the stable node
* is still its ksm page, despite having held no reference to it.
@@ -686,7 +692,8 @@ static void remove_node_from_stable_tree(struct stable_node *stable_node)
* a page to put something that might look like our key in page->mapping.
* is on its way to being freed; but it is an anomaly to bear in mind.
*/
-static struct page *get_ksm_page(struct stable_node *stable_node, bool lock_it)
+static struct page *get_ksm_page(struct stable_node *stable_node,
+ enum get_ksm_page_flags flags)
{
struct page *page;
void *expected_mapping;
@@ -728,8 +735,15 @@ static struct page *get_ksm_page(struct stable_node *stable_node, bool lock_it)
goto stale;
}
- if (lock_it) {
+ if (flags == GET_KSM_PAGE_TRYLOCK) {
+ if (!trylock_page(page)) {
+ put_page(page);
+ return ERR_PTR(-EBUSY);
+ }
+ } else if (flags == GET_KSM_PAGE_LOCK)
lock_page(page);
+
+ if (flags != GET_KSM_PAGE_NOLOCK) {
if (READ_ONCE(page->mapping) != expected_mapping) {
unlock_page(page);
put_page(page);
@@ -763,7 +777,7 @@ static void remove_rmap_item_from_tree(struct rmap_item *rmap_item)
struct page *page;
stable_node = rmap_item->head;
- page = get_ksm_page(stable_node, true);
+ page = get_ksm_page(stable_node, GET_KSM_PAGE_LOCK);
if (!page)
goto out;
@@ -863,7 +877,7 @@ static int remove_stable_node(struct stable_node *stable_node)
struct page *page;
int err;
- page = get_ksm_page(stable_node, true);
+ page = get_ksm_page(stable_node, GET_KSM_PAGE_LOCK);
if (!page) {
/*
* get_ksm_page did remove_node_from_stable_tree itself.
@@ -1385,7 +1399,7 @@ static struct page *stable_node_dup(struct stable_node **_stable_node_dup,
* stable_node parameter itself will be freed from
* under us if it returns NULL.
*/
- _tree_page = get_ksm_page(dup, false);
+ _tree_page = get_ksm_page(dup, GET_KSM_PAGE_NOLOCK);
if (!_tree_page)
continue;
nr += 1;
@@ -1508,7 +1522,7 @@ static struct page *__stable_node_chain(struct stable_node **_stable_node_dup,
if (!is_stable_node_chain(stable_node)) {
if (is_page_sharing_candidate(stable_node)) {
*_stable_node_dup = stable_node;
- return get_ksm_page(stable_node, false);
+ return get_ksm_page(stable_node, GET_KSM_PAGE_NOLOCK);
}
/*
* _stable_node_dup set to NULL means the stable_node
@@ -1613,7 +1627,8 @@ static struct page *stable_tree_search(struct page *page)
* wrprotected at all times. Any will work
* fine to continue the walk.
*/
- tree_page = get_ksm_page(stable_node_any, false);
+ tree_page = get_ksm_page(stable_node_any,
+ GET_KSM_PAGE_NOLOCK);
}
VM_BUG_ON(!stable_node_dup ^ !!stable_node_any);
if (!tree_page) {
@@ -1673,7 +1688,12 @@ static struct page *stable_tree_search(struct page *page)
* It would be more elegant to return stable_node
* than kpage, but that involves more changes.
*/
- tree_page = get_ksm_page(stable_node_dup, true);
+ tree_page = get_ksm_page(stable_node_dup,
+ GET_KSM_PAGE_TRYLOCK);
+
+ if (PTR_ERR(tree_page) == -EBUSY)
+ return ERR_PTR(-EBUSY);

or just:

if (PTR_ERR(tree_page) == -EBUSY)
return tree_page;

right?

+
if (unlikely(!tree_page))
/*
* The tree may have been rebalanced,
@@ -1842,7 +1862,8 @@ static struct stable_node *stable_tree_insert(struct page *kpage)
* wrprotected at all times. Any will work
* fine to continue the walk.
*/
- tree_page = get_ksm_page(stable_node_any, false);
+ tree_page = get_ksm_page(stable_node_any,
+ GET_KSM_PAGE_NOLOCK);
}
VM_BUG_ON(!stable_node_dup ^ !!stable_node_any);
if (!tree_page) {
@@ -2060,6 +2081,10 @@ static void cmp_and_merge_page(struct page *page, struct rmap_item *rmap_item)
/* We first start with searching the page inside the stable tree */
kpage = stable_tree_search(page);
+
+ if (PTR_ERR(kpage) == -EBUSY)
+ return;
+
if (kpage == page && rmap_item->head == stable_node) {
put_page(kpage);
return;
@@ -2242,7 +2267,8 @@ static struct rmap_item *scan_get_next_rmap_item(struct page **page)
list_for_each_entry_safe(stable_node, next,
&migrate_nodes, list) {
- page = get_ksm_page(stable_node, false);
+ page = get_ksm_page(stable_node,
+ GET_KSM_PAGE_NOLOCK);
if (page)
put_page(page);
cond_resched();


Hi Yang,

The patch looks correct as far doing what it claims to do. I'll leave it
to others to decide if a trylock-based approach is really what you want,
for KSM scans. It seems reasonable from my very limited knowledge of
KSM: there shouldn't be any cases where you really *need* to wait for
a page lock, because the whole system is really sort of an optimization
anyway.


thanks,
--
John Hubbard
NVIDIA